alinurdin/time-series-benchmark

Four-Domain Real-World Forecasting Benchmark

Release r2026_10, schema 1.0, data version 2026-09-26.4, evaluator 0.5. What changed since r2026_09 is listed under "Changes in r2026_10" below.

This is a small evaluation corpus for general-purpose time-series forecasting models. It contains real observations in four domains: electricity load, road traffic, river discharge and near-surface weather. There are 13 datasets, 1,049 entities, 23.0M rows and 43.2M measurement slots (41.5M non-null values), about 180 MB of Parquet in total. Two columns hold operator day-ahead forecasts; they are not counted as measurements.

Eleven datasets come straight from the original provider. Two are harmonised redistributions: OPSD (upstream ENTSO-E and European operators) and Caravan (upstream national hydrometric agencies). Native sampling is kept. Nothing is resampled, interpolated or normalised, except one Caravan unit conversion noted below.

Datasets

Dataset	Domain	Provider	Entities (groups)	Where	Period	Time step
eia_hourly	electricity	US EIA, Hourly Electric Grid Monitor (EIA-930)	48 balancing authorities	US: 16 southeast, 16 west, 6 southwest, 4 central, 2 northeast, 2 midwest, 1 Texas, 1 mid-Atlantic	2018–2025	1 h
aemo_dispatch_5min	electricity	AEMO, MMSDM DISPATCHREGIONSUM	5 NEM regions	NSW, QLD, SA, TAS, VIC	2018–2025	5 min
opsd_load	electricity	Open Power System Data, release 2020-10-06	16 areas	AT, LU (15 min); GB, IE (30 min); BG, CH, DK, EE, FI, IT, ME, NO, PL, PT, RS, SK (1 h)	Jan 2015–Jun 2020	per area
madrid_15min	traffic	Madrid City Council, Calle 30 (M-30) ring road	24 detectors (8 groups of 3; 6 are co-located lanes)	Madrid, 4 ring quadrants	2023–2024	15 min
webtris_15min	traffic	National Highways WebTRIS	24 sites (8 corridors of 3)	England: M1, A1(M), M4, M6, M25, M27, M32, M60	2023–2024	15 min
usgs_daily	streamflow	USGS Water Data for the Nation	192 gauges (158 holdout groups)	Conterminous US, 8 regions × 24	2010–2024	daily
usgs_continuous	streamflow	USGS Water Data for the Nation	24 gauges (12 pairs: 11 neighbouring-basin, 1 upstream/downstream)	US, 6 regions: 4 pairs Pacific NW/California, 3 Southwest–Great Basin, 2 Missouri–Arkansas, 1 each Northeast, South Atlantic–Gulf, Great Lakes–Ohio	2021–2024	irregular; median 15 min (21 gauges), 5 min (2) or 10 min (1)
hydat_daily	streamflow	Water Survey of Canada HYDAT	192 gauges (159 holdout groups)	Canada, 6 settings × 32, all provinces/territories except PEI	2010–2024	daily
bom_daily	streamflow	Bureau of Meteorology Water Data Online	128 gauges	Australia, 4 regimes	2010–2024	daily
caravan_daily	streamflow	Caravan v1.6	240 catchments	60 each: CAMELS-BR, CAMELS-CL, CAMELS-GB, LamaH (Austria, Germany, Switzerland)	1999–2013	daily
isd_hourly	weather	NOAA Integrated Surface Database	48 stations (24 groups)	8 each in Africa, Asia, Europe, North America, South America, Oceania	2021–2024	irregular, routine ~hourly reports
swissmetnet_10min	weather	MeteoSwiss SwissMetNet	12 stations (4 transects of 3)	Switzerland: 3 north-Alpine, 1 south-Alpine transect	2023–2024	10 min
ghcnd_daily	weather	NOAA GHCN-Daily	96 stations	23 Oceania, 23 North America, 22 Europe, 15 Africa, 13 Asia; no South America	2015–2024	daily

Panels were selected deterministically, using only static metadata and fitting-era data. When a stratum had too few eligible entities, its unfilled slots went to other strata and the shortfall is recorded in selection/<dataset_id>.json. As a result some panels are less even than their design:

• eia_hourly: five regions have fewer than their six planned BAs.
• bom_daily: 27 tropical monsoonal gauges instead of 32.
• ghcnd_daily: no South American stations (none reaches 70% Tmax coverage in 2015–2021; GHCN-Daily's South American maximum temperatures are sparse), and 13 rather than 16 in Asia (two Saudi stations were selected in the Africa stratum and are now labelled Asia).

Variables

Dataset	Column	Unit	Meaning	Non-null
eia_hourly	demand	MWh	Demand served over the preceding hour (equal to average MW)	99.7%
	demand_forecast	MWh	The BA's own day-ahead forecast (EIA-930 DF). Issued forecast; reference baseline only	99.3%
aemo_dispatch_5min	total_demand	MW	Regional dispatch demand, non-intervention run. Not AEMO's operational-demand product	100%
opsd_load	load	MW	Actual load over the preceding interval; each area's step is in entities.cadence_iso	99.7%
	load_forecast	MW	TSO day-ahead load forecast (ENTSO-E). Issued forecast; reference baseline only	99.1%
madrid_15min	speed	m/s	Mean detector speed	93.7%
	flow	1/h	Flow as a rate (vehicles/hour), not a count	100%
	occupancy	1	Occupied fraction of the interval	97.6%
webtris_15min	speed	m/s	Mean vehicle speed; null when no vehicle passed	99.6%
	vehicle_count	1	Vehicles counted in the preceding 15 min (int64)	100%
usgs_daily	discharge	m³/s	Daily mean, converted once from ft³/s	100%
usgs_continuous	discharge	m³/s	Instantaneous, converted once from ft³/s	100%
hydat_daily	discharge	m³/s	Daily mean	100%
bom_daily	discharge	m³/s	Provider daily mean	100%
caravan_daily	discharge	m³/s	Converted once from Caravan's mm/day using the catchment area (area_km2)	100%
isd_hourly	air_temperature, dew_point	°C	Screen-level, instantaneous	99.7%, 99.2%
	station_pressure	hPa	Station-level pressure (ISD MA1)	71.7%
	sea_level_pressure	hPa	Sea-level pressure (ISD SLP); never mixed with station pressure	62.9%
	wind_speed, wind_direction	m/s, °	Direction is null for calm/variable winds	99.5%, 91.7%
swissmetnet_10min	air_temperature, dew_point	°C	2 m, instantaneous	100%
	station_pressure	hPa	At station height, never reduced to sea level	100%
	wind_speed, wind_direction	m/s, °	10-minute means	99.8%
	precipitation	mm	10-minute total; zeros are real	83.2%
ghcnd_daily	air_temperature_max, air_temperature_min	°C	Daily extremes	97.9%, 97.5%
	precipitation	mm	Daily total; trace amounts are flagged T, not stored as exact zeros	83.4%

"Non-null" is the share of stored rows that carry a value. Rows are only stored where the provider reported something, so it does not measure completeness against a regular grid.

Quality columns:

Column	Datasets	Content
discharge__qc	usgs_daily, usgs_continuous	Approval status and qualifier, joined by | (Approved or Provisional)
discharge__qc	hydat_daily	HYDAT symbol: E estimated, A partial day, B ice, D dry, R revised
discharge__qc	bom_daily	BoM KiWIS quality code: 10 A best available, 90 B compromised, 110 C estimate, 140 E unknown
total_demand__qc	aemo_dispatch_5min	N non-intervention only, I an intervention run was also published
demand__qc, demand_forecast__qc	eia_hourly	Release plausibility code: N passes, L zero or negative, H above 2x the BA's fitting-era 99th percentile, R unchanged for 24 h
load__qc	opsd_load	Same codes as EIA
flow__qc, flow__coverage	madrid_15min	Record status: provider error flag (N, or B where the provider left it blank, mostly zero-flow outage placeholders), or Z zero flow for 24 h and R an identical record for 4 h (failed detector; a missing channel counts as unchanged); and the integration samples over the detector's complete count (15 at most detectors, 5 at six). Both apply to the whole record
<variable>__qc	isd_hourly	ISD quality code. For wind direction the wind-type code is appended (C calm, V variable, ...). station_pressure__qc is the release code X where the value is more than 80 hPa from the standard atmosphere at the station elevation
<variable>__qc	ghcnd_daily	GHCN measurement, quality and source flags joined by |

The EIA, OPSD and Madrid codes are computed by the release from causal rules: each looks only at the value itself, at a threshold fitted on the fitting era, or backwards in time, so the same code can mask a model input. caravan_daily, webtris_15min and swissmetnet_10min have no QC columns.

Layout and conventions

benchmark.json                           catalog: domains, weights, 17 tasks, splits, metrics
datasets/<dataset_id>/dataset.json       provenance, column semantics, limitations, file SHA-256s
datasets/<dataset_id>/entities.parquet   one row per entity
datasets/<dataset_id>/observations/part-00000.parquet
selection/<dataset_id>.json              frozen panel, descriptors, quota shortfalls, reserves

Every dataset has one observations file with these columns:

• entity_id (string)
• time
• one nullable column per variable: float64, except vehicle_count, which is int64
• optional __qc (string) and __coverage (float32) columns

The key (entity_id, time) is unique and rows are sorted by it. Missing observations are simply absent rows or nulls; no gap is filled.

time comes in two kinds:

• timestamp[us, tz=UTC] (EIA, AEMO, OPSD, Madrid, WebTRIS, SwissMetNet): the end of the aggregation interval. ISD and USGS continuous are also UTC timestamps, but they are the actual report instants and do not follow a regular grid.
• date32 (USGS daily, HYDAT, BoM, Caravan, GHCN): the provider's own date label. Day boundaries are provider-defined (BoM's retained series is a 24-hour day in local standard time, stamped 00:00) and are not mapped to UTC.

entities.parquet always contains the following columns:

• entity_id, name, country, region, timezone
• latitude, longitude (null for BAs, NEM regions and OPSD areas), elevation_m
• split_group, which keeps physically linked entities together in holdouts
• coverage_start, coverage_end

It also has dataset-specific columns, such as:

• drainage_area_km2, huc, regulated (river gauges)
• rhbn (HYDAT reference basin)
• response_class (USGS continuous)
• road, direction, chainage, travel_direction, corridor_position (WebTRIS)
• quadrant (Madrid)
• cadence_iso (OPSD)
• station_kind (ISD)

Loading

import pandas as pd

repo = "hf://datasets/alinurdin/time-series-benchmark"
obs = pd.read_parquet(f"{repo}/datasets/eia_hourly/observations/")
entities = pd.read_parquet(f"{repo}/datasets/eia_hourly/entities.parquet")

from huggingface_hub import snapshot_download

path = snapshot_download("alinurdin/time-series-benchmark", repo_type="dataset", revision="r2026_10")

datasets.load_dataset("alinurdin/time-series-benchmark", "<dataset_id>") also works, but it loads observations only.

Tasks

benchmark.json defines 17 tasks. Each task lists:

• its targets and permitted past covariates
• known-future calendar features
• context length, scored horizons and the rolling-origin step
• fit/validation/test boundaries as explicit UTC instants (dates for daily data), with the first and last origin of every block
• a quality_policy naming which QC codes are withheld from inputs and scoring
• for the irregular ISD and USGS continuous data, a target_schedule that fixes what each forecast step means
• the unseen-entity and unseen-region holdouts

benchmark.json also carries a protocol block with the boundary, origin and schedule rules.

Task	Targets (past covariates)	Context	Horizons	Test targets
eia_hourly_demand	demand	28 d	1, 6, 24 h	2023–2025
aemo_dispatch_total_demand	total_demand	7 d	30 min, 6 h, 24 h	2023–2025
opsd_load_actual	load	28 d	1, 6, 24 h	2019–Jun 2020
madrid_speed	speed (flow, occupancy)	7 d	15, 30, 60 min, 3 h	2024
madrid_joint	speed, flow, occupancy	7 d	15, 30, 60 min	2024
webtris_speed	speed (vehicle_count)	7 d	15, 30, 60 min, 3 h	2024
webtris_joint	speed, vehicle_count	7 d	15, 30, 60 min	2024
usgs_daily_discharge, hydat_daily_discharge, bom_daily_discharge	discharge	365 d	1, 3, 7 d	2021–2024
caravan_daily_discharge	discharge	365 d	1, 3, 7 d	2010–2013
usgs_continuous_discharge	discharge	7 d	1, 6, 24 h	2023–2024
isd_station_state	air_temperature, station_pressure, wind_speed (dew_point, sea_level_pressure, wind_direction)	72 h	1, 6, 24 h	2024
swissmetnet_station_state	air_temperature, station_pressure, wind_speed (dew_point, wind_direction, precipitation)	24 h	1, 6, 24 h	2024
swissmetnet_precipitation	precipitation (temperature, dew point, wind speed)	24 h	1, 6 h	2024
ghcnd_daily_temperature	Tmax, Tmin (precipitation)	730 d	1, 3, 7 d	2023–2024
ghcnd_daily_precipitation	precipitation (Tmax, Tmin)	730 d	1, 3, 7 d	2023–2024

Tasks are scored under information tracks, which are loader rules over the same tasks:

• H (headline): target history plus deterministic calendar and solar features.
• M: H plus the listed past covariates and static attributes.
• S: M plus past observations of the task's neighbours. Available for the Madrid, WebTRIS, USGS continuous, ISD and SwissMetNet state tasks.

M and S scores are reported as paired deltas against the same model's H score. EIA and OPSD tasks also name a reference_forecast column, the operator's day-ahead forecast. It is a baseline, never a model input, because its issue time is not recorded. It is stored as published, provider defects included: PJM's and CISO's EIA forecasts run one hour early in parts of the test period, PJM's 2021 values are unusable, and FPC's carry a level error of about 22%. The dataset manifests list these.

Headline weights:

• Each domain counts 25%.
• Within a domain, the weight is split equally over source families, then products. USGS daily and continuous share one USGS budget; Madrid and WebTRIS get 12.5% each.
• The two precipitation tasks carry zero headline weight and are reported separately.
• Scores never scale with row count.

MAE is the primary metric, with RMSE/MASE secondary and CRPS/WIS for probabilistic forecasts. MAPE is excluded. Inputs follow the track: target history and the listed calendar or solar features in H, plus the listed channels and static attributes in M, plus neighbours' past values in S. Realized future weather or neighbour values, and statistics fitted on validation/test data, count as leakage.

Limitations

• Retrospective snapshot. Values reflect the provider's state at retrieval time (September 2026). Provisional USGS values and GHCN near-real-time reports may have been revised since. No as-of vintages are stored.
• Pretraining exposure. Nothing here proves that a model has never seen these series. OPSD and Caravan are frozen packages with no fresh-holdout path. WebTRIS paused live publication in September 2026.
• Small spatial panels. The Madrid and WebTRIS corridors, Swiss transects and USGS pairs are small by design. Confidence intervals should resample whole split_groups and multi-week time blocks.
• Madrid replaces Caltrans PeMS, which was planned but could not be accessed. The archive publishes no carriageway direction. Six of the eight detector groups are lanes at one site, so their neighbour context is cross-lane, not propagation. The provider does not say which end of the interval it labels, so stored times take the label as the start and shift it to the interval end. If the label is really the end, values sit one step late, never early. Zero speed with zero flow is stored as null speed (no vehicles), not as a standstill.
• USGS continuous pairs are mostly not upstream/downstream. Eleven of twelve are two gauges on neighbouring streams in one six-digit basin, not taken from a provider flow network. One pair (Rock River, Iowa) is a genuine upstream/downstream pair. There is no common cadence; the task scores fifteen-minute bins, taking the last report in each bin.
• Provider errors are flagged, not removed. EIA-930 contains sentinel-like spikes (PJM 2,147,480,000 MWh; TVA about 9.9 million MWh twice in 2024), large negative values and multi-day stuck values; OPSD has two SK spikes, one in the test period. Madrid detector 6659 publishes zero flow from late October 2023 to the end of 2024. ISD's Kano station publishes a "station pressure" about 100 hPa too high for its elevation. Tasks withhold these through their QC codes. Positive dips are not flagged: OPSD GB has 102 test intervals below half its fitting-era 1st percentile (for example 863 MW against a typical 37 GW).
• Caravan streamflow is quantised at 0.01 mm/day in the CSV release, which is about 540 m³/s at the Amazon gauge, so low flows in large catchments form long flat runs.
• Provider estimates are scored. HYDAT ice-period values (26%, symbol B) and other provider estimates are kept and reported as a separate stratum. They can interpolate towards later measurements.
• ISD and USGS continuous carry no timezone, so their tasks use UTC calendar features plus solar geometry.
• ISD reports are irregular. About 9% of hourly bins hold more than one report (special reports, or a METAR shortly before a SYNOP). The task scores the last report per hour. The ISD candidate pool also required stations to survive to the end of 2024, which the other panels avoid.
• ISD pressure is sparse outside Europe, Japan, North America and parts of Oceania. METAR reports there carry no pressure, so pressure exists only at SYNOP hours, and four Australian automatic stations report none at all.
• Daily-hydrology regulation is only flagged for HYDAT (66 of 192 regulated, 4 with no record and left null), not for USGS or BoM. It is the provider's current classification, not a history. The HYDAT panel was chosen under an earlier, incorrect flag, so its regulated/unregulated mix is incidental.
• Local-time sources. Records at ambiguous or non-existent DST local times in Madrid and WebTRIS were dropped rather than reassigned.

Changes in r2026_10

• Fixes:

  • HYDAT regulated flag: 66 of 192, not 6. Gauges without a regulation record are null, not false.
  • ISD panel reselected without Swiss stations (SwissMetNet covers Switzerland), with Middle Eastern and North African stations on the right continent.
  • ISD reports that share a timestamp are resolved per variable. A SYNOP's pressure is no longer discarded behind a METAR at the same time.
  • ISO country codes for ISD and GHCN; raw NOAA codes kept in isd_country / fips_country.
  • Madrid timestamps shifted forward one interval, from the provider label to the interval end.

• New columns:

  • demand_forecast (EIA) and load_forecast (OPSD)
  • ISD station_pressure, now the pressure target
  • HYDAT rhbn
  • WebTRIS travel_direction and corridor_position

• Relabelled: USGS continuous pairs are neighbouring basins; pair_role is removed.

• Catalog: information tracks H/M/S; neighbours for Madrid, WebTRIS and USGS continuous; reference-forecast baselines; UTC calendar plus solar features for ISD and USGS continuous.

• Unchanged: the BoM, Caravan and SwissMetNet datasets are copied from r2026_09 unchanged (only their manifest text is updated). USGS daily was also copied, then reselected in the second audit below; the other panels are unchanged apart from ISD.

• Audit fixes (data version 2026-09-26.2, evaluator 0.3):

  • USGS continuous: the last day of every month was missing, because each request window ended at midnight on its last day. Re-extracted complete.
  • EIA demand__qc / demand_forecast__qc and OPSD load__qc: impossible provider values (sentinels, negative demand, stuck values) are flagged and withheld by the tasks.
  • Madrid flow__qc: failed-detector records (Z, R) are flagged and withheld.
  • Madrid flow__coverage: taken against each detector's complete sample count. Six detectors (two groups) complete a record with 5 samples, not 15, and the old fixed denominator made the tasks' coverage gate withhold all of their data.
  • GHCN: the provider's -999.9 elevation sentinel (one station) is now null.
  • AEMO timezone: each region's civil zone, not the fixed market clock.
  • Catalog: explicit UTC split boundaries and origin schedules, target schedules for ISD and USGS continuous, QC policies for every scored target, and no "unseen region" for single-region AEMO.

• Second audit (data version 2026-09-26.3, evaluator 0.4):

  • USGS daily and continuous panels reselected. The gauge inventory kept only exact eight-digit hydrologic unit codes, but the service returns twelve-digit codes for most states, so the earlier panels came from 8,735 of about 72,500 gauges, with none in the Pacific Northwest/California. usgs_daily now has 24 gauges in each of the eight regions, and usgs_continuous all twelve pairs across six regions.
  • WebTRIS: an average speed of 0 with no vehicles counted (5,353 rows) is stored as null speed, not a standstill.
  • ISD: station pressure inconsistent with the station elevation is coded X and withheld (5,863 values at Kano, 1,931 of them previously scored).
  • USGS continuous targets are fifteen-minute bins (last report in the bin), so the ten-minute gauge is fully scored.
  • Catalog: unseen_region_entity_ids keeps physical groups whole in the region holdout (HYDAT 08MH straddles two regions); a zero-scale rule for MASE; per-dataset data_version; notes on calendar features for date-indexed data and on the retained window's final interval.
  • Manifest limitations updated for measured facts (all-zero river series, provider drainage-area errors, gauges without test data, HYDAT publication lag).

• Third audit (data version 2026-09-26.4, evaluator 0.5):

  • Madrid flow__qc: a blank provider error flag (12,139 records, mostly zero-flow outage placeholders) is coded B and withheld; it was null and scored.
  • Madrid flow__qc: the stuck-record rule R treats a missing channel as unchanged. Stuck records often carry no occupancy, so about 9,100 of them (921 in the test year) had stayed N and been scored.
  • ISD manifest limitations: six stations (not four) have no station pressure; part-time reporters (Cozzo Spadaro is daytime-only from February to October) and stuck alpine anemometers are documented.

Licences and attribution

The data is redistributed under each provider's terms. Credit the provider when you use a dataset.

Dataset	Terms
eia_hourly	US Government work, EIA Open Data terms; attribute the U.S. Energy Information Administration
aemo_dispatch_5min	AEMO public data under AEMO's copyright permissions and terms of use; attribute AEMO
opsd_load	Open Power System Data package; upstream ENTSO-E/TSO terms apply; attribute OPSD
madrid_15min	CC BY 4.0; attribute Ayuntamiento de Madrid
webtris_15min	Open Government Licence v3.0; attribute National Highways
usgs_daily, usgs_continuous	US Government work (public domain); attribute the U.S. Geological Survey
hydat_daily	Open Government Licence – Canada; attribute Environment and Climate Change Canada
bom_daily	CC BY (Bureau of Meteorology Water Data Online terms); attribute BoM and contributing agencies
caravan_daily	CC BY 4.0 (Caravan; Kratzert et al., 2023, Scientific Data); upstream agency terms apply
isd_hourly, ghcnd_daily	US Government work (public domain); attribute NOAA NCEI
swissmetnet_10min	MeteoSwiss open data; free use with attribution to MeteoSwiss